[t5/t0/mt5 models] faster/leaner custom layer norm

docs/source/model_doc/t5.mdx

@@ -263,6 +263,11 @@ print(tokenizer.batch_decode(output_sequences, skip_special_tokens=True))
 
 <a id='scripts'></a>
 
+## Performance
+
+If you'd like a faster performance, install [apex](https://github.com/NVIDIA/apex#quick-start) and then the model will automatically use `apex.normalization.FusedRMSNorm` instead of `T5LayerNorm`.
+
+
 ## Example scripts
 
 T5 is supported by several example scripts, both for pre-training and fine-tuning.

src/transformers/models/t5/modeling_t5.py

@@ -237,14 +237,19 @@ def load_tf_weights_in_t5(model, config, tf_checkpoint_path):
 class T5LayerNorm(nn.Module):
     def __init__(self, hidden_size, eps=1e-6):
         """
-        Construct a layernorm module in the T5 style No bias and no subtraction of mean.
+        Construct a layernorm module in the T5 style. No bias and no subtraction of mean.
         """
         super().__init__()
         self.weight = nn.Parameter(torch.ones(hidden_size))
         self.variance_epsilon = eps
 
     def forward(self, hidden_states):
-        # layer norm should always be calculated in float32
+
+        # T5 uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean
+        # Square Layer Normalization https://arxiv.org/abs/1910.07467 thus varience is calculated
+        # w/o mean and there is no bias. Additionally we want to make sure that the accumulation for
+        # half-precision inputs is done in fp32
+
         variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
         hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
 
@@ -255,6 +260,17 @@ def forward(self, hidden_states):
         return self.weight * hidden_states
 
 
+try:
+    from apex.normalization import FusedRMSNorm
+
+    T5LayerNorm = FusedRMSNorm  # noqa
+
+    logger.info("Discovered apex.normalization.FusedRMSNorm - will use it instead of T5LayerNorm")
+except ImportError:
+    # using the normal T5LayerNorm
+    pass
+
+
 class T5DenseReluDense(nn.Module):
     def __init__(self, config):
         super().__init__()